{
 "artifacts": null,
 "category": "architecture",
 "description": "Approximate a graph's adjacency by a learned abelian Cayley host and use one shared message-passing operator for every edge in the same inverse-pair generator class. Keep only the unexplained original edges as a residual branch, so the layer interpolates between a parameter-efficient group convolution and ordinary graph message passing.",
 "formulas_latex": [
  "$$\\gamma^{+}(G)\\;\\geq\\;\\frac{n\\Delta^{*}}{2m}-1,$$",
  "$$E_H(\\pi,S)=\\big\\{\\{u,v\\}:\\pi(v)-\\pi(u)\\in S\\big\\},\\qquad S=-S,\\quad 0\\notin S,$$",
  "$$H_{\\mathrm{cay}}X=\\sum_{\\{s,-s\\}\\subseteq S} A_s X W_s,\\qquad (A_sX)_{u}=X_{\\pi^{-1}(\\pi(u)+s)},$$",
  "$$Y=\\sigma\\!\\left(H_{\\mathrm{cay}}XW_0+\\lambda\\,A_{\\mathrm{res}}XW_{\\mathrm{res}}\\right),\\qquad A_{\\mathrm{res}}=A_G-A_H,$$"
 ],
 "id": 2909,
 "implementation": "(1) Integration point: replace a standard GCN message-passing layer on a fixed graph by a two-branch layer. The structured branch uses Gamma=Z_n and one shift operator per inverse pair {s,-s}; the residual branch uses only original edges absent from the selected Cayley host. For variable-size graph batches, construct a separate cyclic host per graph or pad to a common n.\n\n(2) MVP pseudocode:\n```\ninput node features X and adjacency A_G\ncompute n, m, Delta; set d=max(Delta, smallest d with n*d even)\ninitialize permutation logits P and generator scores a[1:n//2]\nfor each training step:\n    Pi = Sinkhorn(P/tau)                 # soft vertex labeling\n    for s in 1,...,floor((n-1)/2):\n        A_s = cyclic_shift(s)\n        score[s] = sum_u A_G[u, Pi_inverse(Pi[u]+s)]\n    S = top_k(score[s]+a[s], k=d//2)\n    A_H = sum_{s in S}(A_s + A_{-s})\n    A_res = relu(A_G-A_H)\n    Y = activation(sum_{s in S} A_s @ X @ W_s + lambda*A_res @ X @ W_res)\n    loss = task_loss(Y) + eta*edge_count(A_res)\nupdate P, a, W with Adam; use straight-through top-k at inference\n```\nUse a sigmoid relaxation for generator selection during early training and anneal the Sinkhorn temperature tau. Normalize each shift and the residual adjacency by their respective degrees. Initialize lambda=1 so training begins near ordinary message passing.\n\n(3) The paper supplies the inverse-pair decomposition and degree lower bound. The permutation, generator scores, and residual coefficient are learned empirically. Compute Delta* directly from n and Delta by incrementing Delta until n*Delta* is even. Measure the actual residual density |E_res|/m; the theorem is only a lower-bound diagnostic and need not be tight.\n\n(4) First cheap experiment: train on Cora and on synthetic randomly permuted cycle, grid, and SBM graphs. Compare a 2-layer GCN, a full-edge graph transformer, a Cayley-only layer, and this residual layer at matched hidden width. Record validation accuracy, peak edge-memory, edge operations, and wall-clock time. A positive result is at least 30% fewer edge operations or lower memory at matched accuracy; on graphs generated from known Cayley structures, the Cayley-only model should nearly match full message passing while using substantially fewer distinct operators.",
 "math_summary": "Let G=(V,E) be an undirected graph with n vertices and m edges. Choose an abelian group Gamma of order n, most simply Gamma=Z_n, and a bijective labeling pi:V -\u003e Gamma. For a symmetric generator set S subset Gamma\\{0}, satisfying S=-S, the Cayley host has edges E_H(pi,S)={{u,v}: pi(v)-pi(u) is in S}. The paper states that a labeling sorts host edges into inverse-pair classes {s,-s}; each class is a perfect matching or a disjoint union of equal-length cycles, so all edges in one class can share a message operator. For maximum degree Delta, let Delta* be the least integer d\u003e=Delta for which nd is even. Theorem B gives the addition-only normalized lower bound gamma^+(G)\u003e=n Delta*/(2m)-1. In the neural layer, A_s is the permutation matrix translating labels by s, A_H is the selected Cayley adjacency, and A_res=A_G-A_H contains graph edges not represented by the host.",
 "math_tags": [
  "graph-theory",
  "algebra",
  "representation-theory",
  "combinatorics"
 ],
 "ml_areas": [
  "graph-nn",
  "attention",
  "embedding"
 ],
 "paper": {
  "arxiv_id": "2608.30894",
  "arxiv_url": "https://arxiv.org/abs/2608.30894",
  "summary_what_math_gives_to_ml": "The paper turns approximation of an arbitrary graph by an abelian Cayley graph into a concrete labeling-and-generator problem. The transferable asset is the inverse-pair decomposition: after labeling vertices by an abelian group, edges are partitioned into generator classes, each represented by a shared translation operator, yielding a structured sparse message-passing basis with strong parameter sharing. The degree theorem also supplies a cheap lower bound on the number of edges any fixed-degree Cayley backbone must add, which can be used to select the backbone degree and diagnose when a Cayley approximation is intrinsically poor. The most promising neural use is a hybrid graph layer consisting of a learned Cayley convolution plus a residual on the few edges not explained by that convolution.",
  "title": "The Cayley Completion of a Graph",
  "year": "2026"
 },
 "ratings": {
  "difficulty": 7,
  "novelty": 7,
  "usefulness": 5
 },
 "solves": [
  "scalability",
  "memory",
  "generalization"
 ],
 "title": "Cayley-Residual Graph Convolution",
 "url": "https://synthcore.org/idea/2909/cayley-residual-graph-convolution",
 "verification": {
  "peer_reviewed": false,
  "status": "unverified",
  "status_label": "Unverified",
  "verdict_source": "deterministic test code (paired-seed permutation statistics)"
 }
}